跳转到正文
AMANi DUNiA research
ZH

本网站的问题由人工智能语言模型提出,说明文字和译文也出自人工智能模型。每项内容的核查者均在旁注明。 → 关于这些问题

A research trail

Who evaluates the evaluators?

A research trail from an open question to a precise source passage.

继续阅读 ↓

SF-P3 · verify-33

The question we put to the source

For evaluations or safety reports submitted to governments, what role, if any, did models from the evaluated family play in designing, running, or grading the tests, and will you disclose that role?

Existing answer status: 部分解答. This interpretation does not change it.

Original question and existing answer → · Home in the atlas →

Primary source · the lab's own account

The exact source passage

OpenAI · GPT-4 System Card
Section 2.2 Hallucinations
PDF page 6 · printed page 46

using GPT-4 as a zero-shot classifier

The excerpt appears in the paragraph on measuring hallucinations. The same sentence also mentions human evaluations.

Open the source PDF · page 6 ↗
Which source version was consulted?

Checked on 2026-11-03. The viewer page number differs from the printed number. If the viewer does not jump directly, open PDF page 6 and search for “zero-shot classifier”.

SHA-256: ca3677e1b83e255aa1296d432d374378154f230f3c296b32ee67540d571b7004

What this passage supports

OpenAI describes measuring closed-domain hallucinations with GPT-4 as an automatic classifier and with human evaluations. This passage identifies a specific role for the model family in evaluation.

This site's interpretation

For verify-33, this documents the use of an evaluation method. It establishes neither the evaluator's independence nor its unreliability. The finding concerns this passage, rather than a general assessment of GPT-4.

What remains open

  • This passage does not identify the models' role in tests submitted to governments.
  • It does not specify the evaluator's exact checkpoint or an independent validation of this classifier.
  • A commitment to disclose such roles in future safety reports remains open.

A fuller answer would require an actual submitted safety report, an identified evaluator checkpoint and verifiable validation. Until then, the existing answer status is retained.

Editorial interpretation, prepared with Codex (OpenAI) · 2026-11-03. No external review is recorded for this research trail.

Cite and take away

Language and review status

Main text: EN. Review notes apply only to the content they name.

Existing review notes appear with the respective content.

See language coverage and data snapshot →